vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

touchup (+2 squashed commit)

Browse source 
Squashed commit:

[8b06458] fixed broken param order

[7eabdc0] very broken, do not use
This commit is contained in:
Concedo 2023-07-22 22:29:11 +08:00
parent 3aec3038d4
commit 2807d98fd4
3 changed files with 18 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

20
koboldcpp.py
View file

@ -969,7 +969,7 @@ def show_new_gui():
basefile = os.path.basename(model_var.get())
horde_name_var.set(os.path.splitext(basefile)[0])
usehorde_box = makecheckbox(network_tab, "Configure for Horde", usehorde_var, 4, command=togglehorde)
makecheckbox(network_tab, "Configure for Horde", usehorde_var, 4, command=togglehorde)
togglehorde(1,1,1)
# launch
@ -1130,6 +1130,7 @@ def show_new_gui():
if len(dict["hordeconfig"]) > 4:
horde_apikey_var.set(dict["hordeconfig"][3])
horde_workername_var.set(dict["hordeconfig"][4])
usehorde_var.set("1")
def save_config():
file_type = [("KoboldCpp Settings", "*.kcpps")]
@ -1343,7 +1344,7 @@ def show_old_gui():
#A very simple and stripped down embedded horde worker with no dependencies
def run_horde_worker(args, api_key, worker_name):
import urllib.request
global friendlymodelname, maxhordectx, maxhordelen, exitcounter
global friendlymodelname, maxhordectx, maxhordelen, exitcounter, modelbusy
def make_url_request(url, data, method='POST'):
try:
@ -1384,6 +1385,12 @@ def run_horde_worker(args, api_key, worker_name):
break
while exitcounter < 10:
#first, make sure we are not generating
if modelbusy.locked():
time.sleep(0.5)
continue
#pop new request
gen_dict = {
"name": worker_name,
@ -1406,13 +1413,14 @@ def run_horde_worker(args, api_key, worker_name):
continue
current_id = pop['id']
current_payload = pop['payload']
print(f"Job received from {cluster} for {current_payload.get('max_length',80)} tokens and {current_payload.get('max_context_length',1024)} max context. Starting generation...")
print(f"\nJob received from {cluster} for {current_payload.get('max_length',80)} tokens and {current_payload.get('max_context_length',1024)} max context. Starting generation...")
#do gen
while exitcounter < 10:
current_generation = make_url_request(f'http://localhost:{args.port}/api/v1/generate', current_payload)
if current_generation:
break
if not modelbusy.locked():
current_generation = make_url_request(f'http://localhost:{args.port}/api/v1/generate', current_payload)
if current_generation:
break
print("Server Busy - Not ready to generate...")
time.sleep(5)

4
otherarch/gptj_v3.cpp
View file

@ -478,8 +478,8 @@ bool gptj_eval(
// self-attention
{
struct ggml_tensor * Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, freq_base, freq_scale, n_ctx);
struct ggml_tensor * Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, freq_base, freq_scale, n_ctx);
struct ggml_tensor * Qcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, n_ctx, freq_base, freq_scale);
struct ggml_tensor * Kcur = ggml_rope_custom_inplace(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0, n_ctx, freq_base, freq_scale);
// store key and value to memory
{

4
otherarch/neox_v3.cpp
View file

@ -506,8 +506,8 @@ bool gpt_neox_eval(
struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 2*sizeof(float)*n_embd/n_head));
// using mode = 2 for GPT-NeoX mode
Qcur = ggml_rope_custom_inplace(ctx0, Qcur, n_past, n_rot, 2, freq_base, freq_scale, n_ctx);
Kcur = ggml_rope_custom_inplace(ctx0, Kcur, n_past, n_rot, 2, freq_base, freq_scale, n_ctx);
Qcur = ggml_rope_custom_inplace(ctx0, Qcur, n_past, n_rot, 2, n_ctx, freq_base, freq_scale);
Kcur = ggml_rope_custom_inplace(ctx0, Kcur, n_past, n_rot, 2, n_ctx, freq_base, freq_scale);
// store key and value to memory
{